>>43682332See picrel. It depends.
I straight up removed any greyscale/manga/spoken dialogue images from my dataset. Its just polluting it.
If the character names are important to your images then yes. Otherwise I would purge them, they shouldnt be important to the dataset.
>>43682406Automatic aspect bucketing ratio. It auto resizes them into whatever it prefers. You only have to set the average resolution, usually 512 or upto 768. Higher consumes more VRAM which eats into your batch sizes. But can make it more versatile if prompting at a higher base res. Barely matters right now imo. I did it on 512.
